In classical statistics, if your training data suffers from class imbalance, subpopulation shifts, or spurious correlations, the standard remedy is importance weighting. If a minority group is underrepresented by a factor of ten, you simply multiply its loss by ten. This classic tool forces the optimization algorithm to prioritize underperforming samples, shifting the decision boundary to ensure fairness or robustness.
Yet, in modern deep learning, this foundational technique mysteriously breaks down. As models grow large enough to comfortably fit every single training point, multiplying the loss of a sample by a large weight often yields exactly the same decision boundary as standard, unweighted training.
To understand why importance weighting loses its power, we have to look at what happens in the final stages of training when high-capacity models enter the interpolation regime.
The Zero-Gradient Collapse
The most direct intuition for why importance weighting fails in overparameterized networks comes down to a simple algebraic reality: $w \times 0 = 0$.
Consider a standard empirical risk objective where each sample $(x_i, y_i)$ is assigned an importance weight $w_i$:
$$\mathcal{L}(\theta) = \sum_{i=1}^N w_i \ell(f(x_i; \theta), y_i)$$
When we optimize this objective using gradient descent, the parameter updates are driven by the gradient of the loss:
$$\nabla_\theta \mathcal{L}(\theta) = \sum_{i=1}^N w_i \nabla_\theta \ell(f(x_i; \theta), y_i)$$
In classical regimes, models lack the capacity to perfectly satisfy all constraints simultaneously, meaning the individual gradients $\nabla_\theta \ell_i$ remain non-zero, and the weights $w_i$ permanently scale their relative influence.
However, overparameterized neural networks possess enough capacity to achieve near-zero training error across all data subgroups, perfectly memorizing or interpolating every training point. As highlighted by Narasimhan et al., once a model perfectly fits a sample, both its individual loss $\ell_i$ and its gradient $\nabla_\theta \ell_i$ vanish.
When $\nabla_\theta \ell_i$ drops to zero, the importance weight $w_i$ has nothing left to scale. As a result, Mor et al. and Kimura et al. observe that the relative weights assigned to individual samples cease to exert any influence over the final direction of the model's decision boundary. The distributionally robust objective effectively collapses back into standard empirical risk minimization (Sagawa et al.; Zhai et al.).
The Geometry of Implicit Bias
A natural counterargument arises: even if the gradients eventually vanish, shouldn't the importance weights steer the model along a better optimization path before the loss hits zero, leaving a lasting impact on the final decision boundary?
The reason this does not happen is due to the implicit bias of gradient descent. For classification losses with exponential tails (such as cross-entropy), gradient descent does not just stop when the data becomes linearly separable in the feature space; it continues to push the parameters to minimize the loss further toward zero.
As demonstrated by Byrd et al., during this extended training phase, the influence of importance weights steadily diminishes. Asymptotically, the optimization process drives the network toward a specific maximum-margin solution.
Standard Regime (Underparameterized)
[ Minority Class ] <-- (Heavy Weight Shifts Boundary) --> [ Majority Class ]
Interpolation Regime (Overparameterized)
[ Minority Class ] <-- (Boundary Fixed by Data Geometry) --> [ Majority Class ]
As explained by Wang et al., this max-margin destination is dictated entirely by the geometric arrangement of the closest data points (the support vectors) in the feature space, completely independent of the sample weights. Xu et al. clarify this distinction by showing that while importance weights alter the speed of convergence and the generalization error during the early, finite-step phases of training, they cannot alter the ultimate geometric destination of the decision boundary. Once the data is separated, the weights are progressively forgotten.
The Memorization Trap
When we upweight a minority group in a high-capacity model, the network satisfies this heavy penalty not by learning a broad, generalizable rule for that group, but by exploiting its capacity to memorize those specific points.
Zhang et al. show that simply increasing the loss weight for minority samples causes the network to carve out tiny, high-dimensional pockets around those specific training examples to eliminate their loss quickly. Because the model can easily achieve zero training error across all groups via this localized memorization, it defaults back to its original inductive bias for the rest of the feature space.
Consequently, the model continues to rely on easy-to-learn, spurious shortcuts present in the majority data (Raghunathan et al.; Sagawa et al.). This explains why importance weighting fails to protect worst-group accuracy as model scale increases—even up to large language models with hundreds of billions of parameters (Lodkaew et al.). There is a fundamental theoretical limit beyond which loss weighting cannot achieve class or group balance in overparameterized regimes (Stromberg et al.).
Restoring Control Over the Boundary
If passive loss scaling is neutralized by interpolation, how can we force overparameterized models to maintain robust decision boundaries? The literature points to three main paradigms that successfully bypass the vanishing weight problem:
1. Data-Level Interventions (Subsampling)
Instead of upweighting the minority, one can downsample the majority. Sagawa et al. demonstrate that physically removing majority data points completely reverses the harmful effects of overparameterization on worst-group accuracy. Because the max-margin solution is determined purely by data geometry, changing the presence of data points alters the support vectors in a way that scaling their loss cannot.
2. Explicit Regularization and Early Stopping
If the failure of importance weighting stems from the model entering the terminal, zero-error phase of training, one can prevent this collapse by keeping the model away from zero loss. The authors show that pairing importance weighting with early stopping or exceptionally strong regularization (such as heavy weight decay) preserves the influence of the weights by preventing the network from perfectly interpolating the data (Sagawa et al.; Kimura et al.).
3. Geometric and Logit Adjustments
Instead of scaling the loss outside the network's activation functions, active strategies modify the internal geometry of the decision boundary.
- Vector-Scaling (VS) Loss: Behnia et al. and Narasimhan et al. demonstrate that multiplicative logit adjustments apply shifts inside the loss function's tail. These adjustments do not fade during extended training, ensuring a lasting bias toward core features over spurious ones.
- Temperature-Scaled Margins: Martinez et al. introduce frameworks like TempLoss, which decouple the model learner from the fairness objective by using temperature-scaled margins. This ensures that the optimization process remains sensitive to group constraints even after the training data has been fully interpolated.
Key Foundational Work
Overparameterization hurts worst-group accuracy with spurious correlations (2020) — contributed 4 concept elements to this direction.
An Investigation of Why Overparameterization Exacerbates Spurious Correlations (2020) — contributed 3 concept elements to this direction.
What is the Effect of Importance Weighting in Deep Learning? (2019) — contributed 3 concept elements to this direction.